How to issue a single prediction for each area?

[osamamer]

Search before asking

• I have searched the YOLOv5 issues and discussions and found no similar questions.

Question

I have an instance segmentation task, and I am looking for a way to make segment/predict.py produce only one detection for each area. So for instance, if a group of pixels is predicted as both class A and class B, I want it to only give me the class prediction with the higher confidence.
I have tried setting multi_label to false in utils/general.py but that hasn't worked.

Additional

No response

[github-actions]

👋 Hello @osamamer, thank you for your interest in YOLOv5 🚀! Please visit our ⭐️ Tutorials to get started, where you can find quickstart guides for simple tasks like Custom Data Training all the way to advanced concepts like Hyperparameter Evolution.

If this is a 🐛 Bug Report, please provide a minimum reproducible example to help us debug it.

If this is a custom training ❓ Question, please provide as much information as possible, including dataset image examples and training logs, and verify you are following our Tips for Best Training Results.

Requirements

Python>=3.8.0 with all requirements.txt installed including PyTorch>=1.8. To get started:

git clone https://github.com/ultralytics/yolov5  # clone
cd yolov5
pip install -r requirements.txt  # install

Environments

YOLOv5 may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

If this badge is green, all YOLOv5 GitHub Actions Continuous Integration (CI) tests are currently passing. CI tests verify correct operation of YOLOv5 training, validation, inference, export and benchmarks on macOS, Windows, and Ubuntu every 24 hours and on every commit.

Introducing YOLOv8 🚀

We're excited to announce the launch of our latest state-of-the-art (SOTA) object detection model for 2023 - YOLOv8 🚀!

Designed to be fast, accurate, and easy to use, YOLOv8 is an ideal choice for a wide range of object detection, image segmentation and image classification tasks. With YOLOv8, you'll be able to quickly and accurately detect objects in real-time, streamline your workflows, and achieve new levels of accuracy in your projects.

Check out our YOLOv8 Docs for details and get started with:

pip install ultralytics

[glenn-jocher]

@osamamer hello! For your instance segmentation task where you want a single prediction per area, you can adjust the non-maximum suppression (NMS) settings to filter overlapping detections. By default, YOLOv5 applies NMS to each detection, but if you're seeing multiple labels for the same area, you might need to fine-tune the NMS iou (intersection over union) threshold.

To ensure only the highest confidence class is retained for a given area, you can:

1. Increase the iou_thresh parameter in the NMS function to be more strict about overlapping detections.
2. Ensure multi_label is set to False to get only the top class per detection.

You can find these settings in the detect.py script or wherever you're calling the NMS function. If you need more guidance on how to adjust these parameters, please refer to our documentation.

Remember, the goal is to balance the iou threshold so that it's high enough to merge detections of the same object but not so high that it merges distinct objects. Fine-tuning this parameter may require some experimentation based on your specific dataset.

If you continue to face issues, please provide more details about your current configuration and the results you're getting, and we'll be happy to help you further. Good luck with your segmentation task! 😊

[osamamer]

@osamamer hello! For your instance segmentation task where you want a single prediction per area, you can adjust the non-maximum suppression (NMS) settings to filter overlapping detections. By default, YOLOv5 applies NMS to each detection, but if you're seeing multiple labels for the same area, you might need to fine-tune the NMS iou (intersection over union) threshold.

To ensure only the highest confidence class is retained for a given area, you can:

1. Increase the `iou_thresh` parameter in the NMS function to be more strict about overlapping detections.

2. Ensure `multi_label` is set to `False` to get only the top class per detection.

You can find these settings in the detect.py script or wherever you're calling the NMS function. If you need more guidance on how to adjust these parameters, please refer to our documentation.

Remember, the goal is to balance the iou threshold so that it's high enough to merge detections of the same object but not so high that it merges distinct objects. Fine-tuning this parameter may require some experimentation based on your specific dataset.

If you continue to face issues, please provide more details about your current configuration and the results you're getting, and we'll be happy to help you further. Good luck with your segmentation task! 😊

Thank you very much for the reply!
This has worked but unfortunately now there are correct detections that are no longer present no matter how much I reduce the iou_thres.
My task is instance segmentation of X-ray photos of teeth, and now when there is a clear well-defined root in the midst of a big detection of bone, it no longer predicts it as anything.
I am running the segment/predict.py script on my test images after having trained the model on a training set ten times the size for 500 epochs (it early stopped at around 400).
There is an idea to instead make two binary classifiers for bone and root each because these are the ones that cause all the trouble in the predictions. Is this a good idea? Or is there a better way to fix the predictions with the model we already have?

[glenn-jocher]

@osamamer, thanks for the update and for sharing more details about your task. It sounds like you're encountering a common challenge in instance segmentation, especially with complex images like X-ray photos where distinct objects can be closely intertwined.

Given your situation, there are a couple of strategies you might consider:

1. Adjust the Confidence Threshold: If reducing the iou_thres leads to missing correct detections, you might also want to experiment with adjusting the confidence threshold. Lowering it slightly could allow more detections to be considered before NMS filters them out, though be mindful of increasing false positives.

2. Class-specific NMS: Since your issue involves specific classes (bone and root), you could implement or adjust class-specific NMS thresholds. This way, you can have different iou_thres values for different classes, allowing you to be more lenient or strict depending on the class characteristics.

3. Two-stage Detection: Your idea of using two binary classifiers for bone and root separately is a form of a two-stage detection process. This can be effective, especially if these classes have significantly different characteristics. You would first detect all potential objects with one classifier and then refine the detections with the second classifier focused on the challenging classes. This approach requires more computational resources but can improve accuracy for specific tasks.

4. Post-processing: Sometimes, custom post-processing logic can help. For instance, if you know that roots are always within a certain size range or always appear in relation to bones, you can use this knowledge to adjust detections after the initial prediction.

5. Data Augmentation and Training Adjustments: Finally, if you haven't already, consider revisiting your training data and process. More targeted data augmentation, focusing on challenging scenarios, or even manually annotating more examples of difficult cases can improve model performance. Additionally, ensuring your model is neither underfitting nor overfitting is crucial. Sometimes, training for fewer epochs or on a more curated dataset can yield better real-world performance.

Each of these strategies has its trade-offs and might require some experimentation to find the right balance for your specific task. It's often a combination of these adjustments rather than a single change that leads to the best results.

Keep iterating, and good luck with your project! If you have further questions or need more detailed advice, feel free to ask.

[github-actions]

👋 Hello there! We wanted to give you a friendly reminder that this issue has not had any recent activity and may be closed soon, but don't worry - you can always reopen it if needed. If you still have any questions or concerns, please feel free to let us know how we can help.

For additional resources and information, please see the links below:

• Docs: https://docs.ultralytics.com
• HUB: https://hub.ultralytics.com
• Community: https://community.ultralytics.com

Feel free to inform us of any other issues you discover or feature requests that come to mind in the future. Pull Requests (PRs) are also always welcomed!

Thank you for your contributions to YOLO 🚀 and Vision AI ⭐